Resource Constrained Dialog Policy Learning Via Differentiable Inductive Logic Programming

Proceedings of the 28th International Conference on Computational Linguistics, pages 6775–6787 Barcelona, Spain (Online), December 8-13, 2020 6775 Resource Constrained Dialog Policy Learning via Differentiable Inductive Logic Programming Zhenpeng Zhou1, Ahmad Beirami1, Paul Crook1, Pararth Shah1, Rajen Subba1, and Alborz Geramifard1 1Facebook {zzp, beirami, pacrook, pararths, rasubba, alborzg}@fb.com Abstract Motivated by the needs of resource constrained dialog policy learning, we introduce dialog pol- icy via differentiable inductive logic (DILOG). We explore the tasks of one-shot learning and zero-shot domain transfer with DILOG on SimDial and MultiWoZ. Using a single representative dialog from the restaurant domain, we train DILOG on the SimDial dataset and obtain 99+% in-domain test accuracy. We also show that the trained DILOG zero-shot transfers to all other domains with 99+% accuracy, proving the suitability of DILOG to slot-filling dialogs. We fur- ther extend our study to the MultiWoZ dataset achieving 90+% inform and success metrics. We also observe that these metrics are not capturing some of the shortcomings of DILOG in terms of false positives, prompting us to measure an auxiliary Action F1 score. We show that DILOG is 100x more data efficient than state-of-the-art neural approaches on MultiWoZ while achieving similar performance metrics. We conclude with a discussion on the strengths and weaknesses of DILOG. 1 Introduction To foster research on dialog policy learning for virtual digital assistants, several task-oriented dialog corpora have been introduced in recent years, such as SimDial (Zhao and Eskenazi, 2018), Multi- WoZ (Budzianowski et al., 2018), Taskmaster (Byrne et al., 2019), and Schema Guided Dialog (Rastogi et al., 2019), to name a few. Deep learning approaches, including mixture models (Pei et al., 2019) hi- erarchical encoder/decoder (Zhang et al., 2019; Chen et al., 2019), reinforcement learning (Zhao et al., 2019), and pre-trained language models (Wu et al., 2019; Peng et al., 2020; Hosseini-Asl et al., 2020), have significantly advanced dialog policy research in the past few years (Gao et al., 2019), setting new state-of-the-art performance limits. However, collecting annotated data for supervised dialog policy learning is an expensive and time- consuming process. Hence, it is desirable to explore approaches to train dialog policy with limited data and transfer an existing policy with few or even no additional training data to new domains. This practical requirement has motivated the community to research resource-constrained dialog policy learning in the past few decades. Researchers have explored approaches including employing grammar constraints for dialog policy (Eshghi et al., 2017), transfer learning (Shalyminov et al., 2019), or pre-trained language models (Zhao et al., 2020). Few-shot domain adaptation has been researched since the 2000s (Litman and Pan, 2002) on both end-to-end dialog systems (Qian and Yu, 2019; Zhao and Eskenazi, 2018) as well as dialog policy learning (Vlasov et al., 2018). In a traditional modular dialog system, the dialog policy aims to decide a dialog action given a dialog state, while assuming the tasks of language understanding and generation are handled by other com- ponents. Under such assumptions, a task-oriented dialog policy mostly follow the slot-filling scheme, …

Content truncated to 500 words